Overview

adc_cpp targets a single compute stack (mesh + transport + Poisson + AMR), but this stack can run on six parallel configurations: from single-thread serial up to multi-GPU Grace-Hopper distributed by MPI. The key design point is that no operator changes from one configuration to another: all parallelism is confined to two seams (dispatch seams). The backend is chosen at compile time, through CMake options.

This page describes each configuration: what it is, the build command, how to run it, and its validation status (tested in CI or validated manually on ROMEO). For the test-by-test coverage matrix, see BACKEND_COVERAGE.md (single source of truth); for the phase-by-phase GPU port, see GPU_RUNTIME_PORT.md.